Ten newsrooms that build their own AI tools

From vibe coding in the layout to your own language model

Daniel Schiepe had never built an app before. He is the layout editor for the hi-fi magazine *Stereo* at Falkemedia in Kiel. It’s a problem familiar to many layout designers: for many product reviews, there isn’t time to use the in-house photo studio, so the manufacturer’s press image ends up in the magazine and clashes with its overall look. Since early 2026, an app built by Schiepe himself has been converting such images to match the “Stereo” style. Cases like this are becoming increasingly common in editorial offices. If a suitable tool cannot be found, people simply build one themselves using AI, without the need for IT support or a project plan.

jumping-arrow-white

Published: 19 November 2026 | Photo / Video: Magnific

The Ukrainian journalist Katya Gorchinskaya is collecting in the J·Index Publicly documented AI projects by media organisations. Over 3,000 cases from 142 countries have been compiled. Just under a third come from small, local and hyperlocal newsrooms, whilst around 60 per cent of the implementations relate to the newsroom. Gorchinskaya was asked what sets the organisations that are well ahead apart. Her answer: it is neither money nor size that makes the difference, but a senior management team that is curious to try out AI and takes it seriously as both an opportunity and a threat.

For the German-language book market, the AI study conducted by IG Digital at the Börsenverein in collaboration with the management consultancy Highberg points in the same direction. The proportion of in-house AI developments rose from 24.5 per cent in 2025 to 32.8 per cent this year. One in three publishers is now developing AI in-house. Partnerships with external providers remain stable at around 40 per cent. According to the study, publishers are continuing to purchase AI solutions whilst also developing their own.

From the toolbox

Most of these in-house developments start with standard tools. The trade magazine *kress pro* has documented eleven case studies from the Kiel-based publisher Falkemedia. At ‘Stereo’, it wasn’t just the layout that was improved using a bespoke app; producing the newsletter had long been a time-consuming task there. Articles were selected from the website and copied manually into a rigid tool, a process that took up to five hours. Today, the editorial team works with a project in the AI assistant Claude, which contains style templates and instructions for the newsletter. With a quick command, it finds the new content on stereo.de, sorts them according to editorial criteria, generates the HTML code for the newsletter tool and provides a welcome message relating to that particular issue. The process takes around 15 minutes. For technical reasons, the editorial team still updates music reviews manually, using templates prepared by the system. A development department was not involved at any stage.

‘Sylt Fräulein’, the magazine run by blogger and editor-in-chief Finja Schulze, deals with something more delicate than time: namely, a voice. The AI editorial assistant is based on over 300 published texts from the magazine, website and blog, as well as social media posts and an analysis of community comments. There are specific style guidelines stored for each format. After an appointment, Schulze sends notes, voice recordings and photos directly from her car to the AI assistant. By the time she arrives at her next appointment, a rough draft in the magazine’s characteristic tone is already available, which a colleague in the editorial team checks, supplements and approves. Whether the assistant really writes in the ‘Sylt Fräulein’ style depends, according to Falkemedias’ IT director Sebastian Fischbeck, primarily on the volume of texts used to train it. Ten examples aren’t enough. Only with hundreds of texts, sorted by text type and section, does it eventually hit the right brand tone. If there’s too little material, it produces run-of-the-mill texts that take so long to revise that you might as well have written them yourself in the first place.

Whilst at Falkemedia individual members of the editorial teams build their own tools, the UK arm of the US media group ‘Hearst’ – with magazines such as Cosmopolitan and Esquire – is promoting the development on a broader scale. All teams have access to an enterprise version of ChatGPT and are encouraged to build their own GPTs for routine tasks, such as crafting headlines. However, before the company rolled out AI tools, staff were able to disclose which AI services they had already been using on their own initiative, without fear of reprisal. This was accompanied by training sessions and an AI policy.

Toby Wiseman, Managing Director of Content, draws a clear line, however. AI is not to be allowed to write articles at Hearst. He sees greater potential in the archives of the company’s 17 magazine brands, which have grown over decades and have hardly been utilised to date. AI is to be used to reach readers in a more targeted way, for example through personalised recommendations. According to Wiseman himself, however, these are still at the experimental stage. The aim is to turn them into established processes.

When the editorial team is coding vibe

Back to layout designer Daniel Schiepe. His app was created using the Google AI Studio builder. He drags two images – a cut-out of the device and a reference photo from the ‘Stereo’ studio – into a simple browser interface. The result is a lead image in 4K resolution, over 5,000 pixels wide; the AI handles the shadows and lighting. If the cut-out image is missing, the app generates it at the touch of a button.

However, the AI often distorts logos, lettering and technical details. The editorial team therefore subsequently aligns the generated image precisely with the original cut-out. Editor-in-chief Julian Kienzle describes this process as essential, as a fictional control knob on an amplifier would be disastrous in the test magazine. The team has completely removed some functions. The AI was unable to reliably adjust background colours or reposition devices that had been photographed from a significantly different angle, even with more detailed instructions. The app can therefore do less than planned, but it does so reliably.

Even large organisations build their own systems, using their own teams. At the New York Times, the nine-strong AI Initiatives team launched the ‘News Agent’ in July. This in-house chatbot searches the newspaper’s archive, databases, the organisation’s own style guide and the open web. Each reporter can also use it to set up their own research agent for a specific project. This feature was most recently used primarily for the Epstein files. The US Department of Justice had published more than three million pages of documents relating to the case of convicted sex offender Jeffrey Epstein – more than any one person could ever read. NPR, the US public radio network, has also built its own search tool for the Epstein files. This enables individual departments to find the documents relevant to their topics.

The Wall Street Journal’s data journalism team has built its own analysis tool called ‘Doppler’. It analyses large volumes of video and audio content and categorises it by topic. For its coverage of the US congressional elections in November, Doppler searched through more than 25 million TikTok videos and around 500,000 podcast episodes. This allows the editorial team to see what voters are actually talking about on social media and to turn this into their own stories.

Moj Kraj, a Croatian news portal for local news that is almost entirely powered by AI, takes this a step further. The software developer Tihomir Jauk set it up because many small towns in Croatia now have virtually no local news coverage of their own. AI programmes monitor more than 2,000 sources for over 124 towns and villages – such as local council and authority websites – and use this information to draft news stories. Each draft is checked against set criteria, such as neutrality. Of around 100 drafts produced each day, only 20 to 30 are published.

Will AI replace the editorial team?

Kölner Stadt-Anzeiger Medien has closed the Rheinische Redaktionsgemeinschaft with effect from 30 September 2026. Until now, it had operated as a separate service provider producing the regional sections covering the Cologne area. 57 full-time posts will be lost. Its tasks will be taken over by a fully digital production process, in which recurring tasks are largely automated. Editor-in-chief Gerald Selch promises that the time freed up will be channelled into research, analysis and on-the-ground reporting. As an example, he cites an AI system that will in future be able to analyse council documents from all local authorities in the region. The publisher also points to more than 30 full-time posts that have been created in recent months for journalistic initiatives.

What these tools have in common

Thomson Reuters is one of the world’s largest providers of specialist information, best known for its legal and tax databases such as Westlaw and for the Reuters news agency. At the end of August 2026, the company unveiled ‘Thomson’, its own language model for law and tax. A freely available model served as the basis and was further trained using the company’s own specialist content. According to the company, this currently accounts for less than a tenth of the content. ‘Thomson’ is initially being deployed in CoCounsel, the AI assistant for legal professionals. However, the company’s own model does not completely replace the purchased ones. CoCounsel continues to use AI models from other providers, such as Anthropic. ‘Thomson’ only takes on tasks where, in the company’s assessment, it performs better – initially, the review of large volumes of documents.

So even a company with its own language model continues to rely on models from other providers. This is consistent with what Katya Gorchinskaya has observed in her database. Editorial assistants are very similar across organisations and countries. They draw on their own data, such as archives and editorial guidelines, run on one or more models, and offer functions such as drafting, archive searches or transcription, either as a standalone application or within the editorial system.

The model can be swapped out, as Thomson Reuters demonstrates. Nowadays, anyone who has never programmed before can build the interface. What cannot be bought, however, are the 300 articles from ‘Sylt Fräulein’, the style templates from the ‘Stereo’ newsletter, the New York Times style guide and decades of specialist content. At the same time, Gorchinskaya warns against two forms of dependency. Anyone who bases their work on systems controlled by others relinquishes control. And anyone who outsources too much thinking to machines will, over time, lose their ability to judge.

What does this mean for publishers?

The examples that work have one thing in common: they all began with a task whose outcome could be measured, whether in terms of hours saved or the number of designs. 

Raphael Baier, Managing Director of Fabasoft Xpublisher GmbH, sees this as the first prerequisite for any in-house development:

Anyone who has not defined the target state in measurable terms is effectively perpetuating the status quo and thereby entrenching inefficiencies.

Raphael Baier, Geschäftsführer der Fabasoft Xpublisher GmbH

This gives rise to five points for publishers.

  • Look after your own equipment. An AI tool is only as good as the texts, rules and templates it works with. In the case of ‘Sylt Fräulein’, these were hundreds of brand texts; for the ‘Stereo’ newsletter, they were style guides. Publishers should therefore designate someone within the organisation to manage and keep style guides, archives and sets of rules up to date. This task is just as important as deciding which tool to use.

  • Enable staff to build. Falkemedia simply let the editorial team get on with it. Hearst gave everyone corporate access and declared an amnesty beforehand. Both approaches only work if management explicitly wants to try new things and expects that not everything will go to plan.

  • Build a control mechanism into the system. At ‘Stereo’, every AI-generated image is compared with the original photograph; at ‘Sylt Fräulein’, a colleague approves each draft; at ‘Moj Kraj’, only content that meets the review criteria is published. Every tool needs a stage like this, where a person or a fixed rule catches errors before they are published.

  • Do not become dependent on a single AI provider. Despite having its own model, Thomson Reuters continues to use several models in parallel. This allows the company to switch providers if a supplier changes its terms and conditions or the quality declines.

  • Redefine roles and job descriptions. When a layout designer builds their own apps and, in future, an AI analyses council documents, job profiles will change. Job descriptions and further training should take this into account.

Conclusion

Perhaps as early as next year, a better tool will replace Daniel Schiepe’s app. However, he will still need the reference photo from the ‘Stereo’ studio. It is only this photo – which he still produces himself – that sets out the rules for the AI on what a lead image in ‘Stereo’ should look like. The same applies to all the examples in this text. Even if editorial teams can build their own AI-powered tools, their most important asset remains the creation of their own content. No AI provider can deliver that.

The article is part of the Channels: Digital Publishing Technologies, which focuses on content strategies and processes. The channel is sponsored by Fabasoft Xpublisher.

This magazine uses cookies 🍪

Some cookies are technically necessary for everything to work. Others enable additional content that enhances the user experience. You can optionally allow or block these. More on this in our and imprint .